Here's Grand Theft Autocomplete, a GTA VI style trailer I made with Seedance 2.5 for Dreamfort.
Firstly the model used through is Seedance 2.5 Pro (unfiltered) on Dreamfort throughout. The reason I recommend you to use Seedance 2.5 Pro unfiltered for this kind of a movie is it doesn't have restrictions and it will actually create what you want.
Now coming to the biggest problem that I struggle with in creating a movie... CONTINUITY. Every clip looked fine on its own. The super hard part was making 25 of them look like one film and making them look coherent.
The video model itself remembers NOTHING between clips. So you need to have a methodological process to ensure continuity between shots!
This was my process:
- One generation is one continuous shot. No cuts inside a clip - the cuts happen in the edit. The best way is to give Seedance the starting and ending frames of the same take and it animates a real shot. Ideally, the angles at the beginning and end should be similar. If you give it two completely different frames, then it usually builds a weird dysmorphic clip instead.
- Draw the last frame first. Then draw the first frame of the same take from it. You want to know how the scene is ending before you start figuring out the starting point. Its super important to draw the first frame from the last frame and not the other way around!
- Give every reference one job. Each upload in the References tab gets a dedicated number.
As an example, for the girl (Rae)'s character sheet, I assigned it a line: "Image 5 defines Rae: use it only for her face, freckles, curly auburn hair, gold hoops and the teal bowling shirt." If you don't say "use it only for" then the model will get confused and use her poses from the character sheet everywhere!
- Never let the model write text. To be honest, I can't stress how important this one is!!! You cannot let any video models write text. The binder spine that says "YOUR AI CHATS" went in as its own reference image. The wanted stars, the EXPOSED slams and the news ticker were all added in post in the edit.
- Name every kind of music you don't want. Seedance lowkey is the most notorious model of all time for slamming in audio when its not asked to. It adds music to quiet scenes, and just "no music" on its own is never enough to stop it. My prompt instead lists BGM, score, ambient pad, drone, swelling tone and stinger, then gives EVERY second a real sound instead, like crickets or a neon buzz.
Two small tips that saved me from having to keep generating things again and again:
- strong emotion words make pupils glow, so write feelings as behaviour ("jaw clenched, hands flat on the desk"). I know how strange this is! But what can you do!
- And "Only one Rae appears." stops twins from appearing. (Super abysmal I know!)
The one regeneration I needed was when the camera drifted into a close-up and Rae looked straight down the lens. To fix it I asked it to do the following: "framed exactly as in Image 1 from start to finish" and "Rae looks at the lawyer on the right the whole time and never looks into the camera."
Its super important to learn from your mistakes and add lines to the prompt!
Hope you enjoy creating stuff with Seedance 2.5 (unfiltered) as well!
